md/raid1: make sure resync waits for conflicting writes to complete.
authorNeilBrown <neilb@suse.de>
Wed, 10 Sep 2014 05:01:49 +0000 (15:01 +1000)
committerNeilBrown <neilb@suse.de>
Mon, 22 Sep 2014 01:26:01 +0000 (11:26 +1000)
The resync/recovery process for raid1 was recently changed
so that writes could happen in parallel with resync providing
they were in different regions of the device.

There is a problem though:  While a write request will always
wait for conflicting resync to complete, a resync request
will *not* always wait for conflicting writes to complete.

Two changes are needed to fix this:

1/ raise_barrier (which waits until it is safe to do resync)
   must wait until current_window_requests is zero
2/ wait_battier (which waits at the start of a new write request)
   must update current_window_requests if the request could
   possible conflict with a concurrent resync.

As concurrent writes and resync can lead to data loss,
this patch is suitable for -stable.

Fixes: 79ef3a8aa1cb1523cc231c9a90a278333c21f761
Cc: stable@vger.kernel.org (v3.13+)
Cc: majianpeng <majianpeng@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid1.c

index a31c92bbcfc9ff7f573d19ac8fba9750e76bf2a7..b13c218997da77efde68b9fb4d177eb17c33a070 100644 (file)
@@ -845,10 +845,12 @@ static void raise_barrier(struct r1conf *conf)
         * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
         *    next resync will reach to the window which normal bios are
         *    handling.
+        * D: while there are any active requests in the current window.
         */
        wait_event_lock_irq(conf->wait_barrier,
                            !conf->array_frozen &&
                            conf->barrier < RESYNC_DEPTH &&
+                           conf->current_window_requests == 0 &&
                            (conf->start_next_window >=
                             conf->next_resync + RESYNC_SECTORS),
                            conf->resync_lock);
@@ -915,8 +917,8 @@ static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
        }
 
        if (bio && bio_data_dir(bio) == WRITE) {
-               if (conf->next_resync + NEXT_NORMALIO_DISTANCE
-                   <= bio->bi_iter.bi_sector) {
+               if (bio->bi_iter.bi_sector >=
+                   conf->next_resync) {
                        if (conf->start_next_window == MaxSector)
                                conf->start_next_window =
                                        conf->next_resync +